Deep Learning-Driven Friendly Jamming for Secure Multicarrier ISAC Under Channel Uncertainty

This paper proposes a deep learning-driven framework for secure multicarrier ISAC systems that utilizes radar echo feedback and a novel nonparametric Fisher Information Matrix estimator to optimize directional friendly jamming and beamforming under channel uncertainty and unknown eavesdropper locations, while employing a quantized tensor train encoder for efficient implementation.

The Gist

Imagine a high-tech lighthouse (the Base Station) that does two jobs at once: it sends secret messages to its friends (the Users) and uses a powerful radar beam to scan the surrounding ocean for ships (the Sensing).

The problem? There's a sneaky pirate ship (The Eavesdropper) lurking in the fog, trying to steal the secret messages. In the old days, to stop the pirate, the lighthouse needed to know exactly where the pirate was hiding. But pirates are good at hiding; they don't send signals, so the lighthouse can't see them clearly.

This paper proposes a clever new way to protect the messages using Deep Learning and a concept called "Friendly Jamming." Here is how it works, broken down into simple parts:

1. The "Radar Echo" Trick (No Need to See the Pirate)

Usually, to jam a pirate, you need to know their exact location. But this paper says, "We don't need to see the pirate; we just need to see what the pirate reflects."

• The Analogy: Imagine the lighthouse shines a flashlight into the fog. Even if you can't see the pirate, the light hits the pirate's ship and bounces back as a "ghostly reflection" (a radar echo).
• The Solution: The lighthouse uses these reflections to figure out, "Ah, there's a suspicious shape in that direction!" It doesn't need to know who the pirate is or have a perfect map of them. It just knows, "Something is there, so I should blast noise in that direction."

2. The "Friendly Jamming" (The Noise Cannon)

Once the lighthouse spots a suspicious direction, it fires a special kind of noise called Friendly Jamming.

• The Analogy: Think of it like a DJ at a party. The DJ wants the VIPs (the Users) to hear the music clearly, but the DJ wants to drown out the person trying to record the conversation (the Pirate).
• The Magic: The DJ (the Base Station) is smart enough to aim the noise only at the pirate. The VIPs are standing in a "quiet zone" (a mathematical null space) where the noise doesn't reach them. The pirate, however, gets blasted with static, making their recording useless.

3. The "Uncertainty" Problem (The Fog is Thick)

In the real world, the radar isn't perfect. The "ghostly reflection" might be blurry, or the pirate might be moving fast. This is called Channel Uncertainty.

• The Old Way: Traditional methods would say, "If the map isn't perfect, we can't jam safely," or they would jam everywhere, wasting energy and accidentally bothering the VIPs.
• The New Way (Deep Learning): The lighthouse uses an AI Brain (a Neural Network) that has been trained on thousands of "what-if" scenarios. It learned that even if the radar is blurry, it can still guess the right direction to jam. It's like a seasoned sailor who can navigate by the stars even when the sky is cloudy.

4. The "Cramér-Rao" Rule (Don't Get Too Blurry)

There's a catch: If you jam too hard, you might mess up your own radar ability to see the pirate. You need a balance.

• The Analogy: Imagine you are trying to take a photo of a bird while also trying to scare it away with a loud noise. If you scream too loud, you might startle the bird so much it flies away before you can get a good picture.
• The Solution: The paper introduces a rule called the Cramér-Rao Lower Bound (CRLB). Think of this as a "Quality Control Meter." The AI is programmed to ensure that even while it's screaming (jamming), the photo (radar reading) stays sharp enough to be useful. It finds the perfect volume of noise that scares the pirate but keeps the radar clear.

5. The "Tiny Brain" (Making it Fast and Small)

Deep Learning models are usually huge and slow, like a supercomputer that takes up a whole room. You can't fit that on a standard cell tower.

• The Innovation: The authors used a technique called Quantized Tensor Train (TT-Q).
• The Analogy: Imagine you have a massive encyclopedia (the big AI model). Instead of throwing away the books, you compress them into a tiny, high-tech microchip that holds all the same knowledge but takes up 100 times less space.
• The Result: The system is now small, fast, and cheap enough to run on real-world hardware without slowing everything down.

6. The "Traffic Lane" Strategy (Multicarrier)

The paper also looks at how to handle many different frequencies (like radio channels) at once.

• The Analogy: Imagine a highway.
  • Overlapping: You can have cars (data) and noise (jamming) on the same lane, but you have to be very careful.
  • Non-Overlapping: You can dedicate specific lanes only for cars and other lanes only for noise.
• The Finding: The AI figured out the best mix. Sometimes it's better to mix them, and sometimes it's better to separate them, depending on how much traffic there is and how foggy the radar is.

Summary: Why This Matters

This paper solves a major headache for future 6G networks. It shows how we can:

1. Protect secrets without needing to know exactly where the spy is.
2. Use radar to help protect the communication.
3. Work even when the sensors are imperfect (which they always are in the real world).
4. Run on small, efficient hardware thanks to the new "compressed brain" technology.

In short, it's a smart, self-correcting security system that uses the environment itself to fight off eavesdroppers, ensuring your data stays safe even in a chaotic, noisy world.

Technical Summary

1. Problem Statement

The paper addresses the critical security vulnerabilities in Integrated Sensing and Communication (ISAC) systems. While ISAC improves spectrum efficiency by combining radar sensing and wireless communication, it exposes both user data and environmental sensing information to eavesdroppers (Eves).

Key Challenges:

• Imperfect Channel State Information (CSI): In practical scenarios, the Base Station (BS) cannot obtain perfect CSI for the eavesdropper (who is passive) or even for legitimate users due to noise, quantization, and mobility.
• Unknown Eavesdropper Location: Traditional friendly jamming (FJ) often requires precise knowledge of the Eve's Angle of Arrival (AoA) or CSI, which is unrealistic.
• Dual Uncertainty: The system must handle both CSI errors and AoA estimation errors simultaneously.
• Complexity and Scalability: Existing optimization-based methods (e.g., convex optimization, SDR) are computationally heavy, assume perfect CSI, and struggle with real-time deployment in multicarrier (OFDM) systems.
• Sensing vs. Security Trade-off: Jamming signals can degrade the radar's ability to estimate target angles. The system must ensure jamming does not violate sensing accuracy constraints (Cramér–Rao Lower Bound - CRLB).

2. Methodology

The authors propose a Deep Learning (DL)-driven framework that replaces traditional optimization with an end-to-end neural network approach, specifically designed for multicarrier ISAC.

A. Core Concept: Radar-Echo-Guided Jamming

Instead of relying on Eve's CSI, the system uses radar echo feedback to identify vulnerable spatial regions.

• The BS transmits probing signals (structured OFDM pilots).
• It collects radar echoes to estimate the angular distribution of potential threats.
• Friendly Jamming (FJ) beams are directed toward these angular regions to degrade Eve's reception without interfering with legitimate users (by projecting FJ into the null space of the legitimate channel).

B. Nonparametric Fisher Information Matrix (FIM) Estimation

To handle sensing uncertainty and ensure the jamming does not ruin radar performance, the paper introduces a novel nonparametric FIM estimator based on $f$-divergence.

• Mechanism: Instead of deriving closed-form FIM equations (which require Gaussian assumptions), the system uses a neural discriminator to estimate the divergence between the original radar echo distribution and perturbed distributions (simulating AoA errors).
• CRLB Constraint: The estimated FIM is used to calculate the CRLB. The optimization ensures the CRLB remains below a specific threshold, guaranteeing that the radar can still accurately estimate angles (Azimuth $\theta$ and Elevation $\phi$) even while jamming.

C. Quantized Tensor Train (TT-Q) Encoder

To address the scalability and memory issues of Deep Neural Networks (DNNs) in high-dimensional MIMO-OFDM systems:

• Tensor Train Decomposition: The fully connected (FC) layers of the encoder are factorized into low-rank tensor cores.
• Quantization: The tensor cores are stored using low-bit precision.
• Result: This reduces the model size by >100x (two orders of magnitude) and significantly lowers inference complexity (FLOPs) while maintaining near-baseline performance.

D. System Architectures

The framework supports two multicarrier schemes:

1. Overlapping Scheme: All subcarriers are used jointly for communication, sensing, and jamming.
2. Non-Overlapping Scheme: Subcarriers are partitioned into disjoint sets (some for communication only, others for sensing/jamming) to reduce mutual interference.

3. Key Contributions

1. Learning-Based Secure ISAC Framework: A novel framework that jointly optimizes beamforming and friendly jamming without requiring Eve's CSI, leveraging real-time radar echoes to guide jamming direction.
2. Robustness to Uncertainty: The system explicitly accounts for imperfect CSI and noisy AoA estimates. It uses a data-driven, nonparametric FIM estimator to satisfy CRLB constraints under uncertainty.
3. Lightweight Implementation: Introduction of a Quantized Tensor Train (TT-Q) encoder that compresses the model by >100x, enabling real-time deployment on resource-constrained edge devices.
4. Flexible Multicarrier Design: The framework supports both overlapping and non-overlapping subcarrier allocations, offering a trade-off between spectral efficiency and interference management.
5. Hardware Resilience: The training process incorporates data augmentation to handle hardware impairments like Phase Noise (PN) and I/Q imbalance, ensuring robustness in practical RF environments.

4. Simulation Results

Extensive simulations were conducted using synthetic Rayleigh fading channels and various impairment levels.

• Secrecy Rate: The proposed ISAC_FJ scheme outperforms baselines (AE_FJ, DL_AN, MRT-Equal) by >50% in secrecy rate. At 20 dB SNR, it achieved ~6.9 nats/s/Hz compared to ~4.5 for baselines.
• Robustness to CSI Errors: The system maintains high secrecy rates even with CSI error variances up to $\rho=0.2$, whereas traditional methods degrade significantly.
• Sensing Accuracy: The method successfully maintains CRLB constraints (sensing accuracy) even as the CRLB threshold increases (worsening conditions), showing superior resilience compared to closed-form steering vector methods.
• Multicarrier Advantage: Performance scales with the number of subcarriers ($N$). A 64-subcarrier system achieved ~6.2 nats/s/Hz, significantly outperforming single-carrier systems which struggled to secure the channel.
• Model Efficiency:
  • Parameter Reduction: From ~62,500 (FC) to ~28,300 (Hybrid TT-FC) and ~2,800 (Pure TT).
  • Inference Speed: Training time per epoch reduced from 2.4s (FC) to 0.6s (TT-based) on an NVIDIA A100 GPU.
  • Performance Loss: The Hybrid TT-FC encoder retained >99% of the FC baseline's secrecy rate.
• Hardware Impairments: The proposed robust training strategy showed negligible secrecy rate degradation under severe phase noise, while non-robust baselines failed.

5. Significance and Impact

This paper represents a significant step toward practical, real-world deployment of secure ISAC systems.

• Shift from Optimization to Learning: It moves away from computationally expensive convex optimization (which fails under dynamic conditions) to efficient deep learning that adapts to channel dynamics.
• Solving the "Unknown Eve" Problem: By using radar echoes rather than Eve's CSI, it solves a fundamental limitation in physical-layer security where eavesdroppers are passive and silent.
• Edge-Ready Security: The TT-Q compression makes it feasible to run complex security algorithms on edge devices (e.g., base stations in 6G networks) with limited memory and compute power.
• Holistic Design: It successfully balances the triad of Security (Secrecy Rate), Reliability (BLER), and Sensing Utility (CRLB), proving that security and sensing can be mutually reinforcing rather than conflicting in ISAC systems.